This invention relates to the architecture of computing systems, and in particular to an architecture in which groups of instructions may be executed in parallel, as well as to methods and apparatus for accomplishing that.
A common goal in the design of computer architectures is to increase the speed of execution of a given set of instructions. Many solutions have been proposed for this problem, and these solutions generally can be divided into two groups.
According to a first approach, the speed of execution of individual instructions is increased by using techniques directed to decreasing the time required to execute a group of instructions serially. Such techniques include employing simple fixed-width instructions, pipelined execution units, separate instruction and data caches, increasing the clock rate of the instruction processor, employing a reduced set of instructions, using branch prediction techniques, and the like. As a result it is now possible to reduce the number of clocks to execute an instruction to approximately one. Thus, in these approaches, the instruction execution rate is limited to the clock speed for the system.
To push the limits of instruction execution to higher levels, a second approach is to issue more than one instruction per clock cycle, in other words, to issue instructions in parallel. This allows the instruction execution rate to exceed the clock rate. There are two classical approaches to parallel execution of instructions.
Computing systems that fetch and examine several instructions simultaneously to find parallelism in existing instruction streams to determine if any can be issued together are known as superscaler computing systems. In a conventional superscaler system, a small number of independent instructions are issued in each clock cycle. Techniques are provided, however, to prevent more than one instruction from issuing if the instructions fetched are dependent upon each other or do not meet other special criteria. There is a high hardware overhead associated with this hardware instruction scheduling process. Typical superscaler machines include the Intel i960CA, the IBM RIOS, the Intergraph Clipper C400, the Motorola 88110, the Sun SuperSparc, the Hewlett-Packard PA-RISC 7100, the DEC Alpha, and the Intel Pentium.
Many researchers have proposed techniques for superscaler multiple instruction issue. Agerwala, T., and J. Cocke [1987] "High Performance Reduced Instruction Set Processors," IBM Tech. Rep. (March), proposed this approach and coined the name "superscaler." IBM described a computing system based on these ideas, and now manufactures and sells that machine as the RS/6000 system. This system is capable of issuing up to four instructions per clock and is described in "The IBM RISC System/6000 Processor," IBM J. of Res. & Develop. (January, 1990) 34:1.
The other classical approach to parallel instruction execution is to employ a "wide-word" or "very long instruction word" (VLIW) architecture. A VLIW machine requires a new instruction set architecture with a wide-word format. A VLIW format instruction is a long fixed-width instruction that encodes multiple concurrent operations. VLIW systems use multiple independent functional units. Instead of issuing multiple independent instructions to the units, a VLIW system combines the multiple operations into one very long instruction. For example, in a VLIW system, multiple integer operations, floating point operations, and memory references may be combined in a single "instruction." Each VLIW instruction thus includes a set of fields, each of which is interpreted and supplied to an appropriate functional unit. Although the wide-word instructions are fetched and executed sequentially, because each word controls the entire breadth of the parallel execution hardware, highly parallel operation results. Wide-word machines have the advantage of scheduling parallel operation statically, when the instructions are compiled. The fixed width instruction word and its parallel hardware, however, are designed to fit the maximum parallelism that might be available in the code, and most of the time far less parallelism is available in the code. Thus for much of the execution time, most of the instruction bandwidth and the instruction memory are unused.
There is often a very limited amount of parallelism available in a randomly chosen sequence of instructions, especially if the functional units are pipelined. When the units are pipelined, operations being issued on a given clock cycle cannot depend upon the outcome of any of the previously issued operations already in the pipeline. Thus, to efficiently employ VLIW, many more parallel operations are required than the number of functional units.
Another disadvantage of VLIW architectures which results from the fixed number of slots in the very long instruction word for classes of instructions, is that a typical VLIW instruction will contain information in only a few of its fields. This is inefficient, requiring the system to be designed for a circumstance that occurs only rarely--a fully populated instruction word.
Another disadvantage of VLIW systems is the need to increase the amount of code. Whenever an instruction is not full, the unused functional units translate to wasted bits, no-ops, in the instruction coding. Thus useful memory and/or instruction cache space is filled with useless no-op instructions. In short, VLIW machines tend to be wasteful of memory space and memory bandwidth except for only a very limited class of programs.
The term VLIW was coined by J. A. Fisher and his colleagues in Fisher, J. A., J. R. Ellis, J. C. Ruttenberg, and A. Nicolau [1984], "Parallel Processing: A Smart Compiler and a Dumb Machine," Proc. SIGPLAN Conf. on Compiler Construction (June), Palo Alto, CA, 11-16. Such a machine was commercialized by Multiflow Corporation.
For a more detailed description of both superscaler and VLIW architectures, see Computer Architecture--a Quantitative Approach, John L. Hennessy and David A. Patterson, Morgan Kaufmann Publishers, 1990.